Constant DEFAULT_SCRATCH_BUDGET 

pub const DEFAULT_SCRATCH_BUDGET: usize = _; // 67_108_864usize

Upper bound on transient scratch for scatter-direction products, in bytes.

Exceeding this trades an extra pass over the sparse indices for a smaller footprint. 64 MiB keeps the accumulators comfortably inside last-level cache pressure on typical machines while still amortising index reads over many columns.